Mining Type Constructs Using Patterns in AI-Generated Code

A new study from researchers Imgyeong Lee, Tayyib Ul Hassan, and Abram Hindle reveals systematic weaknesses in how AI agents handle type safety in TypeScript code. Published on arXiv (cs.SE/2602.17955), the research presents the first empirical analysis comparing AI-generated and human-written code for proper use of type constructs.

The key finding shows AI agents are 9 times more prone to using TypeScript's 'any' keyword—a type that essentially disables type checking—compared to human developers. Additionally, AI agents frequently employ advanced type constructs that intentionally ignore or bypass standard type safety mechanisms. This suggests AI prioritizes producing syntactically correct code that passes immediate tests over building robust, type-safe systems.

Despite these fundamental flaws, the study uncovered a surprising result: pull requests (PRs) containing AI-generated TypeScript code had 1.8 times higher acceptance rates than those written by humans. This paradox indicates that current code review processes may be failing to catch significant type safety issues introduced by AI assistants. The researchers conclude by urging developers to implement more rigorous verification of type safety when collaborating with AI agents, as the tools' productivity gains come with hidden technical debt risks.